SpatialPooler with log boosting #545
Conversation
src/htm/algorithms/SpatialPooler.cpp
Outdated
| if(boost < htm::Epsilon) return; //skip for disabled boosting | ||
| output[i] = exp((targetDensity - actualDensity[i]) * boost); //TODO doc this code | ||
| //output[i] = exp((targetDensity - actualDensity[i]) * boost); //exponential boosting, default for Numenta | ||
| output[i] = log(actualDensity[i]) / log(targetDensity); |
There was a problem hiding this comment.
Performance optimizations, which helped when I did this in the columnPooler2 branch.
- Move the
log(targetDensity)to the outermost loop, since it is constant across all of the mini-columns. In this case calling code contains the loop which you should avoid, not this function. - Use
log2instead oflog
There was a problem hiding this comment.
log2 instead of log
done
log(targetDensity) to the outermost loop, since it is constant
the problem here is I still need to keep local inh working, which changes the density for each column
|
Thanks for putting in the effort to evaluate this function. What I would really like to see from your analysis are the activation frequencies of the cells (AKA variable |
Log boost(optimized for global inh for this case)
Exp boost (str = 7.0)
Exp, str=2.0
Boosting none
Target sparsity is |
does not require parameter
compared old/tested arrays, not the new values
now logboost has 96.2%
This reverts commit c473c4d.
breznak
force-pushed
the
log_boosting_fn
branch
from
70062bc
to
8ebb3e0
Compare
instead of current exponential.
Log boosting has been discussed by @ctrl-z-9000-times on the forums and I believe is actually used in #293
I have tried experimenting with it in the SP.
vs